AAI-M flashcards

32 cards. Read the term and the definition here. The deck at the top of this page is optional.

Machine Learning

A subset of AI in which algorithms learn patterns and parameters from data to make predictions or decisions without being explicitly programmed for each rule.

Supervised Learning

A learning approach that trains a model on labeled input-output pairs so it can predict the target for new, unseen inputs.

Overfitting

A modeling failure where the model fits training noise closely, producing low training error but poor performance on new or held-out data.

Inference

The production phase in which a trained model applies its learned parameters to new inputs to generate predictions or outputs.

Hyperparameter

A configuration value set before training, such as learning rate or tree depth, that governs how the learning algorithm behaves rather than being learned from data.

Data Lineage

A record of a dataset's origin, movement, and transformation steps, used for audit, impact analysis, and troubleshooting data issues.

Class Imbalance

A condition where one target class greatly outnumbers another, making accuracy misleading and favoring precision, recall, or F1 for evaluation.

Data Leakage

A modeling error where information from the test set or future data influences training, producing inflated performance estimates that fail in production.

Concept Drift

A post-deployment change in the relationship between inputs and the target variable, degrading model performance even when input distributions look stable.

Data Quality Dimensions

Standard attributes used to assess data fitness, typically accuracy, completeness, consistency, timeliness, validity, and uniqueness.

Cross-Validation

An evaluation technique that partitions data into rotating folds, training on some and validating on others to estimate generalization more reliably than a single split.

Regularization

A technique that penalizes model complexity, such as large coefficients, to reduce overfitting and improve generalization to unseen data.

Confusion Matrix

A table of true positives, true negatives, false positives, and false negatives used to compute classification metrics such as precision and recall.

MLOps

The practice of automating and managing the machine learning lifecycle, including versioning, testing, deployment, monitoring, and retraining of models.

F1 Score

The harmonic mean of precision and recall, giving a single balanced metric when both false positives and false negatives matter.

Retrieval-Augmented Generation

A technique that retrieves relevant documents and inserts them into the model prompt, grounding responses in source content and reducing hallucination without retraining.

Hallucination

Fluent model output that is factually unsupported or fabricated, often mitigated through grounding, citations, and human review.

Context Window

The maximum number of tokens a language model can process and attend to within a single request, bounding prompt plus output length.

Fine-Tuning

Further training of a pretrained model on curated domain examples to adapt its behavior, distinct from prompt engineering or retrieval augmentation.

Prompt Injection

An attack in which untrusted input manipulates a model's instructions, so model output and downstream actions must be treated as untrusted.

Return on Investment

A value measure computed as benefit minus cost divided by cost, used to compare AI initiatives against alternatives and baselines.

Human-in-the-Loop

A design where a person reviews or approves model output before consequential action, appropriate when errors carry high cost or regulatory weight.

Total Cost of Ownership

The full cost of an AI system across its life, including data preparation, compute, integration, monitoring, and ongoing retraining.

Demographic Parity

A fairness criterion requiring equal positive prediction rates across protected groups, regardless of differences in underlying outcome prevalence.

Equalized Odds

A fairness criterion requiring equal true positive and false positive rates across groups, tying fairness to error rates rather than selection rates.

Model Card

A document describing a model's intended use, training data, evaluation results, and limitations to support transparency and informed downstream use.

Red-Teaming

Structured adversarial testing that probes an AI system for harmful, unsafe, or biased behavior before release and informs mitigation.

Data Minimization

A privacy principle limiting collection and retention of personal data to what is necessary for a stated, lawful purpose.

Risk-Based AI Tiering

A governance approach that classifies AI systems by potential impact and applies documentation, oversight, and testing controls proportionate to that risk.

Model Inventory

A maintained register of deployed and in-development models recording purpose, owner, risk tier, and version to support governance and audit.

Canary Release

A deployment pattern that routes a small share of live traffic to a new model version, limiting impact while real-world performance is observed.

Model Monitoring

Ongoing production tracking of data quality, drift, predictive performance, latency, and cost to detect degradation and trigger retraining or rollback.